An act to add Title 15.3 (commencing with Section 3115) to Part 4 of Division 3 of the Civil Code, relating to artificial intelligence.
AB 412 would create a new Civil Code title governing how developers of generative artificial intelligence models document and disclose the copyrighted and pre-1972 sound recordings used in training. The bill defines key terms such as “GenAI,” “developer,” “covered material,” “rights owner,” and “fingerprint,” and requires developers that commercially use or make a GenAI model available in California to document covered materials they know were used to train the model, make reasonable efforts to identify additional covered materials, and identify the rights owner for each documented item.
The bill also requires developers to post website information that allows rights owners to generate compatible fingerprints and to provide an online mechanism for submitting information requests. After a valid request, the developer must assess whether the provided fingerprints are likely present in its dataset and return a list of the requester’s covered materials that were documented or likely used in training. The bill limits requests to one per calendar quarter per model absent new information, allows requests to cover multiple works, and creates a private right of action with statutory damages, injunctive relief, attorneys’ fees, and other remedies for noncompliance. It exempts certain developers, including those trained exclusively on publicly available no-cost data, noncommercial academic or governmental research, models not trained on covered materials, and models trained only on materials the developer owns.
AB 412 would add a new state-level disclosure and access framework to California Civil Code for GenAI training data involving copyrighted works and pre-1972 sound recordings. It would impose recordkeeping, website disclosure, request-response, retention, and enforcement obligations on covered AI developers operating commercially in California or making models available to Californians, while carving out several categories of developers from coverage. The bill would not change federal copyright ownership rules, but it would create a California enforcement mechanism tied to training-data transparency and rights-holder inquiries, with potential civil liability for missed deadlines or failure to respond.
The bill appears to have generally favorable momentum in committee and on the Assembly floor, as reflected by majority “do pass” votes and a 45-16 third-reading vote. The available voting history suggests the measure advanced with support from a majority of legislators, though not unanimously. No committee transcript was provided, so the record here shows support more clearly than detailed debate or negotiated compromise.
The main points of contention are likely to be the compliance burden on AI developers, the feasibility and cost of fingerprint-based dataset searches, and the scope of the private enforcement provisions. Developers may object to the requirement to document training materials, maintain records for years, and respond within a fixed deadline, especially where training data is large, mixed, or obtained from third parties. Rights holders and supporters are likely to favor the bill’s transparency and enforcement tools, while opponents may argue that the bill could be difficult to implement, may expose trade secrets or proprietary information, and could create litigation risk through per-day violations and statutory damages.